Digitising a 93 page Bible Concordance from scans

TL;DR

Objective: Digitize a printed bible concordance for reprinting. Ideally whole task should be done with free tooling. Final Outcome: Very accurate digital transcription of the whole 93 page document.

Background

The task was to take the scans of a 93 page bible concordance, OCR it accurately for the purposes of reprinting. All we had was high resolution scans of all pages.

Challenges to OCR

The text flows sometimes from one line to next, but sometimes even to the previous line. The editors tried to squeeze in as much content as possible into fewest pages, hence this approach. See picture below showing the flows of each line. Some even overflow into the line above. The output text had to be reassembled as single lines, and type set later for printing. This difficult to do with a conventional OCR tool, which just reads each line without understanding if the text is a continuation from the previous line or next line.

Lines overflowing, even upwards sometimes

The printed quality is not perfect – there are some imperfections in the letters, such as the letter t below. Besides, its quite hard to tell the full stops(.) from the commas(,). The difference is tiny, and a human reader needs contextual knowledge to work out if its a full stop or comma.

The pages are not perfectly straight. Some are slightly tilted, as it was placed in the scanner:

Strategy

The initial approach was to use a conventional Optical Character Recognition(OCR) website to digitize the pages. To make it more accurate, a manual task of splitting each column into a separate image was done then each column was straightened such that each line was parallel with the horizontal. This was done with image editing software Photoshop. Each page has 3 columns and there are 93 pages, so in total, 279 images were made. The preprocessing step was thought to help with accuracy, as OCR tools generally perform better with single column of text all nicely aligned.

Single column, straightened

However, the OCR tool output had of errors. Shown below is a sample output:

There were lots of issues with punctuations, missing letters, some weird characters. Also the overflow lines were handled as seen, with no understanding of the context.

Verification tool

First step was to build a verification tool. A bible concordance has some patterns we can look for with a program.

  1. Ensure that the verse references actually exist in the King James Version bible
  2. Ensure that the root word (e.g. ABIDE) actually exists in the verse referred to in the verse reference
  3. Ensure the verse references follow a consistent pattern.

I built the tool to take the output of the OCR tool, and carry out this simple verification. The output of the OCR tool was an ODT file, like a Microsoft Word doc, but in the open document format. I built a web interface to upload the file, and it produces a page which highlights the compliance with the following rules(there were more than the ones listed here):

  1. Every verse reference must have the following format: <Abbreviated_book>. <chapter>:<verse>. For example, Jn. 3:16 or Ps. 23:5. In cases where there are multiple verses, in same chapter, then must be comma separated. e.g. 1 Jn 2:27, 28(note the space after the comma). Multiple references from different books must be separated by a semi colon. e.g. Mt. 23:12; Lk. 14:11
  2. The root word for each entry in the concordance must be found in the verse the reference points to. e.g. the word ABIDE must exist in Prov. 7:11.
  3. The verse reference, e.g. Ps. 61:7 must exist in the King James bible
  4. All book abbreviations must follow a set list only. e.g. John must be Jn.

If any of the above conditions are not met, the tool flags it with colour coding, and puts up a summary of issues to fix on the left bar. Early on, there were over 10K suspected wrong verse reference, due to issues with exactly matching the required verse reference schema. For example, some didn’t have the blank space between the book and verse, e.g. Lk.13:3 or the fullstop after the book abbreviation was missing e.g. 1 Jn 2:27.

The flow for using this was to upload the hand edited ODT file to the site, and out comes the HTML highlighted version:

These above rules, written in code, were a super tight constraint applied, which meant the final text is very uniform in consistency. The ODT file was gradually hand edited to bring down the large number of mistakes in the transcription(i.e. non compliance with the above rules) but there were simply too many to fix by hand in a reasonable time period. There were over 10k mistakes detected by the verification tool.

I used mostly Claude Sonnet 4.5 and 4.6 to build this tool. It generates very nice looking UIs.

Better Base Line with AI OCR

It became clear that the manual hand editing was taking too long. It was 1 month before letters A to D were complete, and had to be done all the way to letter Z. So I thought it will be a better idea to get a much higher accuracy baseline OCR of the text. The next obvious solution was to use the multimodal AI tools to do the reading.

Qwen 3.5-Plus to digitise

The first attempt to use use AI OCR tool was done with Qwen 3.5-Plus. Initially the results looked impressive. I could use a single prompt and attach up to 4 images(i.e. 4 columns) and it was able to output plausible looking text. The prompt I used and refined was this:

Digitise this image. Do not use web references. Preserve the formatting faithfully. I need maximum accuracy.

There are no greek letters like α in the image. Its more likely to be a. Only normal alphabets so ensure output uses only standard characters. The root words, in uppercase, are to be bolded as is done in the image. If root word not found, which was defined in previous page, then just transcribe the lines.

However, while plausible looking, in some cases, when there were OCR errors, the model was very eager to fill in with plausible looking text that was entirely made up, and not in the original image. Identifying and correcting this was a pain. For example, see the following side by side comparison of the original vs the Qwen 3.5-Plus OCR. See that it completely goes off piste:

In the original it says “God hath given us an u., 1 John 5. 20.” but the digitized output says “Go thou and give them u., 1 John 5. 20.”. These were rare, but in the thousands of lines, it was too many to detect and fix by hand.

DeepSeek-V3.2

Upon hearing feedback that the Qwen output was unworkable, I set about looking for something better. I eventually found DeepSeek-V3.2 to give me good results. At time of writing, I am not 100% sure which model was available on the DeepSeek webchat interface, but I believe it was DeepSeek-V3.2. The accuracy was really good, minimal hallucination. The amazing thing is it was able to reassemble lines that overflowed into the previous line or next line very reliably.

Scanned on left, DeepSeek output on right. Highlighted by colour to show lines that flow to previous or next lines.

I started with single prompt + attached images in a single conversation. DeepSeek was really useful because I could attach up to 50 images and it was very willing to read all of them. So I could do several pages at a time. The path that I took is as follows:

Upload single image with 3 columns(i.e. 1 page worth) stacked vertically into a single image. At this time, I didn’t know that DeepSeek web chat interface accepted many images at a time, so thinking that its limited like Qwen, I uploaded a single image combining 3 columns vertically. My prompt:

Digitise this image. Do not use web references. Preserve the formatting faithfully. I need maximum accuracy.
There are only english characters. The root words, in uppercase, are to be bolded. The abbreviated letter(e.g. a. for APPLE) are in italics
Show each subtopic(under root word) on a separate line.
At the top of the page, the root word may not be found as it was on the previous page. Don’t miss those lines.

Ignore all rate limits. Digitise absolutely every line in the image.

The text coming out was looking good and reasonably accurate. So I carried on, this time uploaded 2 columns as 2 images, with prompt “do the same for the 2 attached images. Must be done by file name sequence”. I had numbered each filename with schema page_<page_number>_1.jpg. This worked well, but I spotted an inconsistency – the abbreviated letter was being bolded. I wanted italics instead. So I steered it with another prompt: “same for this file, ensure abbreviated root word is _italics_”.

And this time, the abbreviated letters were appearing as Italics – exactly as I wanted. Now that enough steering instructions were already in the context, I could do more pages quickly. Gradually I gradually got more and more daring, doing 23 columns at a time. It was working well. When I did this in March 2026, it accepted up to 50 attachments(images) in the prompt. This was much higher than any of the other AI tools.

A few hours later I had all 93 pages of the transcription. I was able to do several pages per prompt. There were occasionally some inconsistencies. For example, sometimes the abbreviation letter was no longer in italics. Its as if the tool forgot about it, so I had to remind it in a prompt. See below:

I then combined all the outputs into a single file. In total, there were 30393 lines, 4529 root words.

Now back to verification

I now had a 30k line document with reasonably accurate transcription, at least when doing spot checks and comparing the original doc with the digital copy. But still with my verification tool, it had around 610 verse references that had issues; the format of the verse reference could be incorrect with missing . or spaces or spelling issues.

Using Claude to fix many of them in a feedback loop

My solution was to get Claude Sonnet 4.6 to fix them as far as possible. I fed the tool with:

  1. The new more accurate odt text file from the Deepseek transcription
  2. The verification tool code
  3. Zip file of the json formatted King James Version bible – required for validation tool

I got Claude to read all the errors and automatically fix the transcript. 800+ errors down to 79.

Some of the fixes that Claude Sonnet 4.6 did was to handle more edge cases in the validation script, and many of the fixes were also by the tool editing the text to add the required dot, colon, space to make the validation pass. The key thing is the tool was able to run the validation script, fix issues, then re-verify in a feedback loop. Allowing the tool to work within this tight feedback loop was an important strategy. It allowed the tool to read the outputs, make corrections, and in each step, optimise towards a numerical target. The numerical target was to bring down the number of incorrect verse references. Claude had all the required information to be able to auto fix – this included the actual King James Bible verses to search through, and the verification script source code, to understand the required schema and also the ability to execute the script and check.

One might say its would have been much easier to use a coding tool like Claude Code to automatically read the files and do the necessary work without me having to manually upload it to the web UI. Its true, but I was looking to do it for free and Claude Code was not free, while the web interface was.

Final edit by hand

The last 100 reference errors were edited by hand. This was very much manageable in this small number, down from over 10k when in the very first OCR of the pages.

Final proof reading by hand

Proof reading by hand, comparing original to the digital text was an essential step. While the verification script checked verse reference format, the short description on each line was not checked. It had to be verified manually by comparing original with digitised copy. So while the AI tools were used to do the grunt work of OCR, and even some validation of schema and correctness was carried out by tool, it was necessary to get the assurance of a real person to read every line and compare with the original printed version. All artifacts of the tooling such as hallucinations by the AI tool could then be caught and eliminated.

What was important to make it a success

Image preprocessing

The input images had to be straightened and split by columns. It was far more accurate this way than to give it the 2 page spread scan, consisting of 6 side by side columns. The AI OCR tools were getting very confused when the original layout was passed in.

Feedback loop

Getting the AI tool to optimise towards a number, that is, reducing the number of incorrect verse references, is really effective.

Verification by machine

The machine is very consistent in its checks, allowing for no deviations. Each verse reference had to follow the exact format, otherwise gets flagged. The final output is a very uniform, consistently formatted document.

This document was entirely written by hand by me, Medad Newman, on 8 – 14 Aug 2026. No Generative AI was used to write it. The project ran from Jan 2026 to June 2026, during which the tooling available changed and improved a lot.

Leave a Reply

Your email address will not be published. Required fields are marked *